Hermite Functions

March 2, 2025

Exercise 5.6, Introduction to Fourier Theory, Stein & Shakarchi.

(i) Define Hm (x) := exp(2πx2 )Dm (exp(−2πx2 )). This is an m-th degree polynomial. Let
Fm (x) := Hm (x) exp(−πx2 ) = exp(πx2 )Dm (exp(−2πx2 )). Then Fm ∈ S(R), since Fm ∈
(l)
C ∞ (R) and ∀k, l, |x|k Fm (x) is bounded by an k + l + m-th degree polynomial in |x| times
exp(−πx ), which decays to 0 as |x| → ∞ since exp(−πx2 ) = o(exp(−π|x|)) = o(|x|k ) for
2

any k > 0.

(ii) Since S(R) is a vector space and S(R) ⊂ M(R) ⊂ L2 (R), equip S(R) with the usual inner
product on L2 (R). Then we show Fm ⊥Fn ∀n ̸= m. Assume w.l.o.g. m > n.

Z
⟨Fn , Fm ⟩ = Fn Fm dλ
ZR
= Hn (x) exp(−πx2 )Hm (x) exp(−πx2 )dx
R
Z
= Hn (x) Hm (x) exp(−2πx2 ) dx
ZR
| {z }
2
= Hn (x)Dm (e−2πx )dx
R
Z
(i) (ii)
= Dm (Hn )(x)dx = 0
R

where (i) holds due to partial integration (m times), which can be applied on the unbounded
domain of integration (without boundary) due to the decay of the integrand as |x| → ∞.
(ii) holds since m > n, and Hn is an n-degree polynomial.
(iii) {Fn }∞ ∞
n=0 is orthogonal; let {Gn }n=0 denote the normalized Fn . Then we show it is maximal.
Since H0 , H1 , . . . are a Hamel basis of the space of all polynomials (Hn has degree n), it
suffices to show the polynomials times exp(−πx2 ) are dense in S(R). Equivalentl is to show
that for any f ∈ S(R), ∀k ∈ Z≥0 : f ⊥xk exp(−πx2 ) =⇒ f ≡ 0.
n R
As follows: if this holds, then F (z) := n≥0 zn! R f (x)xn exp(−πx2 )dx = n≥0 0 = 0 for
P P
n
all z ∈ C. If we fix any compact domain Ω ⊂ C, then the series n≥0 zn! f (x)xn exp(−πx2 )
P

converges uniformly in x ∈ Ω to f (x) exp(−πx2 + zx). Take Ω = B(0, N ). f ∈ S(R) can

be bounded for |d| ≥ N if N large enough and exp(−πx2 ) decays faster than the error of
approximation outside Ω for N large enough, we can first pick N , then conclude uniform
convergence outside and inside Ω, hence on C and thus for x ∈ R. Then we may exchange

1
 integral and series (∗) to obtain:
Z X n Z
(∗) z k 2
F (z) = f (x)x exp(−πx )dx = f (x) exp(−πx2 + zx)dx.
R n! R
n≥0

2
Take z = 2πit, then F (z) is the Fourier transform of f (x) exp(−πx2 ): F (2πit) = F{f (x)e−πx }(t) =
2
fˆ(t) · e−πt . This was 0 for any t, so fˆ ≡ 0, so by Fourier inversion (since f ∈ S(R)), f ≡ 0.
(iv)
X X
Fn (x)tn /n! = exp(πx2 ) Dn (exp(−2πx2 ))tn /n!
n≥0 n≥0

= exp(πx ) exp(−2π(x + t)2 )

2
(1)
2 2
= exp(−πx − 4πxt − 2πt )
= exp(−π(x + 2t)2 + 4πt2 − 2πt2 )
= exp(−π(x + 2t)2 + 2πt2 )

F : S(R) → S(R) is defined from an integral on an unbounded domain R, so in order to

P the series and F we need one of dominated/monotone/uniform convergence of the
exchange
series n≥0 Fn (x)tn /n! as a function of x with t fixed. Note that this series is identically
−πx2
Hn (x)tn /n!. Observe that the leading term of Hn is (−2πx)n , so
P
equal to n≥0 e
n
Hn (x) ∼ (−2πx) as n → ∞, and this holds uniformly in x ∈ C since the non-leading
terms divided by (−2πx)n decay faster if x is larger! And (−2πx)n = o(n!), so taking any
|t| < 1 suffices to have uniform convergence on a compact domain Ω. then use the factor
2
e−πx to uniformly bound the (polynomial) error outside Ω and conclude the convergence
is uniform in x for fixed |t| < 1.
X
F(Fn )(ξ)tn /n! = F{exp(−π(x + 2t)2 + 2πt2 )}(ξ)
n≥0

= exp(2πt2 )F{exp(−π(x + 2t)2 )}(ξ) Linearity

2 2
= exp(2πt )F{exp(−πx )} exp(2πi(2t)ξ) Proposition 1.2
2
= exp(2πt2 ) exp(−πξ 2 ) exp(4πitξ) −πx2 }(ξ) = e−πξ
{e\
= exp(2πt2 − πξ 2 + 4πitξ)
= exp(πξ 2 ) exp(2π(t + iξ)2 )
= exp(πξ 2 ) exp(−2π(ξ − it)2 )
X
= Fn (ξ)(−it)n /n! c.f. (1)
n≥0

For fixed ξ, Hn (ξ) has leading term (−4πξ)n = o(n!) for ξ uniformly in a fixed compact
domain, we have this equality at least for |t| ≤ 1. Convergence and equality of the series
on the unit disk means equality as holomorphic functions in t (for fixed ξ), therefore we
conclude Fn (ξ) = (−i)n Fcn (ξ) needs to hold for all n. So Fn is an eigenvector of F with
n
eigenvalue (−i) .
L L
We now jump to conclusions. n≥0 Fn := n≥0 spanC {Fn } = S(R) (we take the closure
in S(R), which I am aware is incomplete in L2 -norm), hence each eigenspace E(−i)k ⊂ S(R)

2
is of the form
∞
M
E(−i)k = F4n+k
n=0
L
And another application of n≥0 Fn = S(R) allows to conclude that

3
M
S(R) = E(−i)k
k=0

And so the eigenvalues are contained in {±i, ±1}. Each eigenspace is per construction
nomempty as well, so these are the eigenvalues.